Related Experiment Video
Updated: Oct 20, 2025
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
2.1K
Automatic detection of actionable radiology reports using bidirectional encoder representations from transformers.
Yuta Nakamura1,2, Shouhei Hanaoka3,4, Yukihiro Nomura5
1Division of Radiology and Biomedical Engineering, Graduate School of Medicine, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo, 113-8655, Japan. yutanakamura-tky@umin.ac.jp.
BMC Medical Informatics and Decision Making
|September 13, 2021
Summary
Bidirectional encoder representations from transformers (BERT) effectively identify actionable radiology reports. Integrating order information did not significantly improve BERT
Area of Science:
- Medical Informatics
- Natural Language Processing
- Radiology
Background:
- Reliable communication of actionable findings from radiologists to clinicians is crucial.
- Natural Language Processing (NLP) aids in identifying actionable findings in free-text radiology reports.
- Deep learning techniques and order information have not been fully explored for this task.
Purpose of the Study:
- To evaluate the effectiveness of Bidirectional Encoder Representations from Transformers (BERT) for identifying actionable radiology reports.
- To assess the added value of incorporating clinical order information alongside radiology reports.
Main Methods:
- A binary classification task was performed on 90,923 Japanese radiology reports to distinguish actionable from non-actionable reports.
- Four methods were evaluated: Logistic Regression (LR), Gradient Boosting Decision Tree (GBDT), Bidirectional Long Short-Term Memory (LSTM), and a Japanese BERT model.
- Performance was compared using radiology reports alone versus combined with order information.
Main Results:
- The BERT model achieved the highest Area Under the Precision-Recall Curve (AUPRC) and Area Under the Receiver Operating Characteristic Curve (AUROC) when using radiology reports alone.
- BERT significantly outperformed LR, GBDT, and LSTM.
- Incorporating order information yielded only a marginal, statistically insignificant improvement in BERT's performance.
Conclusions:
- Bidirectional Encoder Representations from Transformers (BERT) demonstrate significant utility in detecting actionable radiology reports.
- Further research into more sophisticated methods is needed to effectively leverage clinical order information for improved report classification.
